AI Output Quality Mirrors Your Operating Discipline
The quality of your AI output is a mirror of your operating discipline. Are you using a model, or are you outsourcing requirements?
In a mid-market distributor adding a customer portal on top of SAP, the team asked an AI assistant to “build authentication” and “add an API.” It shipped something impressive and wrong: extra layers, new tables, clever abstractions, and a security story nobody could explain.
The problem was not the model. The problem was the brief.
When you say “make it scalable,” the assistant hears “design a framework.” When you say “keep it minimal,” but never define minimal, you get a small architecture, not a small solution. And if you do not say what not to touch, it will happily rewrite the parts that were stable.
The teams that get leverage treat prompts like a spec: state the target behavior, the constraints (latency, compliance, who owns the data), and the boundaries (do not add new services, do not change Salesforce or ServiceNow integration, one file if possible). Then use a stronger model for the plan and a cheaper one for execution, like you would with humans.
The tell was simple: after tightening prompts, rework dropped, and the review moved from “why did we build this” to “does this meet policy and edge cases.”